The estimation on how well the linear filter in the AEC3 is performing
is done through an estimation of the ERLE. That estimation is then
used for knowing how much the suppressor needs to react in order to
cancel all the echoes.
In the current code, the ERLE is quite conservative during farend
inactivity and it is common that it goes to a minimum value during
those periods. Under highly varying conditions, that is probably the
right approach. However, in other scenarios where conditions does not
change that fast there is a loss in transparency that could be avoided
by means of a different ERLE estimation.
In the current CL, the ERLE estimation has been changed in the
following way:
- During farend activity the ERLE is estimated through a 1st order AR
smoother. This smoother goes faster toward lower ERLE values than to
larger ones in order to avoid overestimation of this
value. Furthermore, during the beginning of the farend burst, an
estimation of the ERLE is done that aim to represent the performance
of the linear filter during onsets. Under highly variant environments,
those quantities, the ERLE during onsets and the one computed during
the whole farend duration, would differ a lot. If the environment is
more stationary, those quantities would be much more similar.
- During nearend activity the ERLE estimation is decreased toward a
value of the ERLE during onsets.
Bug: webrtc:9040
Change-Id: Ieab86370a4333d2d0cd7041047d29651de4f6827
Reviewed-on: https://webrtc-review.googlesource.com/62342
Commit-Queue: Jesus de Vicente Pena <devicentepena@webrtc.org>
Reviewed-by: Per Åhgren <peah@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#22568}
7682c6e2cb